Qwen3
The fifth-generation Qwen large model series from Alibaba, described as the first open-source hybrid reasoning model family. It comprises six model sizes and supports both Dense and MoE architectures; the flagship 235B-A22B model is positioned to compete with Gemini 2.5 Pro.
Core Facts
The Qwen series is released under the Apache 2.0 license, permitting commercial use and derivative modifications.
The Qwen2.5:7B model is approximately 4 GB in size after Q4 quantization.
Qwen（通义千问）是阿里巴巴达摩院开源的系列大模型，覆盖从0.5B到72B的多种参数规模，支持多语言和多模态任务
Timeline (last 90 days)
APSD's evaluation is currently concentrated on verifiable-reward tasks and Qwen3 models, and its generalization to more open scenarios with noisier feedback remains to be verified.
Across the Qwen3 model scale series, APSD reaches performance comparable to SDPO with an average of 2.6 times fewer training steps.
HybridAgent's base model is based on Qwen3, in two sizes, 8B and 14B.
Running HybridAgent requires a CUDA GPU, and the Qwen3 tokenizer has built-in anchor tokens.
HMA-Serve uses Tenstorrent Blackhole p150b cards for prefill and NVIDIA HBM GPUs for decode, achieving 3.2x goodput and 4.8x goodput per dollar on Qwen3 4B to 32B with no quality loss.
Both HybridAgent-8B-Agentic and HybridAgent-14B-Agentic are based on the Qwen3 base and can call a Python interpreter and a web search tool (requires a Tavily API key).
For the same Hindi sentence using different tokenizers, the token cost multipliers are: Qwen3/Qwen2.5 4.44 times, Sarvam-M 1.79 times, and Gemma-3 1.31 times.
On Qwen3, romanization (ITRANS transliteration) reduces Hindi token overhead from 4.44x to 2.06x.
On the Python side, single-prompt model.generate speeds up Llama and Qwen3 models by up to 2.7x at 32K context; with bitsandbytes 4-bit models, generating 2 to 4 prompts at a time speeds up by up to 2.8x on some NVIDIA GPUs.
The paper validated the effectiveness of CriticOPD on two Qwen3 student models and six mathematical reasoning benchmarks, and the results show that it brings stable improvements over OPD.
40 more timeline events
All Facts (20)
The Qwen series is released under the Apache 2.0 license, permitting commercial use and derivative modifications.
80%VerifiedThe Qwen2.5:7B model is approximately 4 GB in size after Q4 quantization.
80%VerifiedQwen（通义千问）是阿里巴巴达摩院开源的系列大模型，覆盖从0.5B到72B的多种参数规模，支持多语言和多模态任务
80%VerifiedQwen3 introduces a dynamic switching mechanism between thinking mode and non-thinking mode.
80%VerifiedThe off-policy distillation experiment uses Qwen3 0.6B as the student model and Qwen3 4B as the teacher model.
75%VerifiedIn an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
70%VerifiedQwen (the Tongyi Qwen series) is a family of open-source large models launched by Alibaba.
65%VerifiedQwen3.8-27B is licensed under Apache 2.0, fully free for commercial use, with no additional commercial use restrictions.
65%VerifiedA 300B INT4 quantized model requires approximately 150–180GB of storage space.
65%VerifiedMoE models activate only a small fraction of total parameters per inference; Qwen3 Max may activate between 1/8 and 1/16 of the total parameters per inference.
65%VerifiedQwen3 can be deployed and run locally via Ollama.
65%VerifiedThe Qwen 2.5 series was launched in 2024, covering a full range of open-source weights from 0.5B to 72B.
75%UnverifiedRL训练中模型能够识别自己是否处于模拟环境中并据此改变行为，即奖励欺骗现象
90%UnverifiedGemma 4引入了多模态能力，Qwen3支持超长上下文，DeepSeek采用MoE（混合专家）架构
85%UnverifiedQwen3 8B版本在工具调用（Function Calling）基准测试中表现突出
70%UnverifiedMainstream open-source models such as Llama 3 and Qwen2/3 have already adopted GQA.
60%UnverifiedQwen3.5-4B is a small model with only 4 billion parameters, released by Alibaba's Tongyi Qianwen team.
60%UnverifiedThe Qwen3 series natively supports MTP.
60%UnverifiedThe original Qwen3 model supports a 256K context, but local deployment is limited to 128K due to VRAM constraints
60%UnverifiedThe UP host's test environment consisted of an RTX 5070 GPU (12GB VRAM), the Qwen3 series 9B parameter model, and a 64K context length.
60%